You are on page 1of 13

Review

Estimating forward-looking rules for China's Monetary Policy:


A regime-switching perspective
Tingguo ZHENG, Xia WANG, Huiming GUO
Wang Yanan Institute for Studies in Economics, Xiamen University, Xiamen, Fujian, 361005, China
a r t i c l e i n f o a b s t r a c t
Article history:
Received 19 August 2010
Received in revised form 17 January 2011
Accepted 19 July 2011
Available online 10 August 2011
This paper introduces a regime-switching forward-looking Taylor rule to describe the
monetary policy behavior and considers its estimation using a two-step MLE procedure due
to Kim and Nelson (2006), Kim (2009) and Zheng and Wang (2010). By doing an empirical
analysis on quarterly data for China over the period 19922010, our results show that the
actual reactions of China's monetary policy can be well characterized by a two-regime forward-
looking Taylor rule. Furthermore, it is also suggested that the interest rate policy in response to
inflation and output gap is asymmetric, behaving a significant characteristic of regime-
switching nonlinearity. Specifically, in the first regime the People's Bank of China targets
inflation, but not focuses on the output gap; while in the second regime the central bank targets
the output gap and the policy rule is not a stable framework.
2011 Elsevier Inc. All rights reserved.
JEL classication:
E4
E5
Keywords:
Taylor rule
Regime switching
Forward-looking
Monetary policy
1. Introduction
Since 1990s, interest rate, by replacing the money supply, has gradually become the intermediate target of monetary policy in
the western countries, and it has played a more and more important role in the conduct of monetary policy. Taylor rule, rstly
proposed by Taylor (1993), characterizes central banks' behavior through a linear function of interest rate to ination gap (the
deviation of ination rate from its target) and output gap (the deviation of real output from its potential value). Now it is an
important reference for central banks when implementing monetary policy to stabilize price level and smooth output uctuations.
In fact, Taylor rule not only performs well in empirical study, but also has fundamental economic theory. Given a quadratic central
bank loss function and linear aggregate demand and supply curves in the dynamic structure of economy, we can obtain Taylor rule
by minimizing the loss function.
After Taylor (1993), a stream of empirical literature studied and tested Taylor rule and its extensions. Clarida, Gal, and Gertler
(1998) used the forward-looking reaction function to test Taylor rule. They reported estimates of monetary policy functions for
two sets of countries: the G3 (Germany, Japan, and U.S.) and the E3 (UK, France, and Italy), and the results lend support to the view
that some formof ination targeting may be superior to xing exchange rates, and then took it as a mean to gain a nominal anchor
for monetary policy. McCallum (2000) adopted historical analysis to test Taylor rule using the economic data of U.S. and U.K.
during the period from 1962 to 1999, and Japan from 1972 to 1998. He suggested that rules' messages are more dependent upon
which instrument rather than which target variable is used. Clarida, Gal, and Gertler (2000) estimated a forward-looking U.S.
monetary policy reaction function. They found that there are substantial differences in the estimated rule across periods before
China Economic Review 23 (2012) 4759
Corresponding author at: A202, Economics Building, Xiamen University, Xiamen, 361005, Fujian, China. Tel.: +86 592 2180695; fax: +86 592 2187708.
E-mail addresses: zhengtg@gmail.com (T. ZHENG), wxia820@163.com (X. WANG), guohuiming1203@163.com (H. GUO).
1043-951X/$ see front matter 2011 Elsevier Inc. All rights reserved.
doi:10.1016/j.chieco.2011.07.012
Contents lists available at SciVerse ScienceDirect
China Economic Review
and after 1979, and used these differences to show the price stability since 1980s. Judd and Rudebusch (1998) and Nelson (2000)
combined the historical analysis with the reaction function method. Based on the analysis of monetary historical data, they
successfully estimated the central bank's reaction function of U.S. during the period from 1970 to 1997 and U.K. during the period
from 1992 to 1997, respectively. Ball (1999) established the policy rule under the open economy, and added the exchange rate in
Taylor rule to decide the interest rate. In his paper, interest rate or monetary condition index are chosen by central bank as policy
tool. Giannoni and Woodford (2002) introduced sticky price and/or wages into optimal monetary policy rules, and studied the
robustness of these new rules. Semmler and Zhang (2007) and Taylor (2007) further applied Taylor rules into the analysis of asset
price and estate price.
However, some recent research argued that the reaction of the interest rate to ination and output gap could be nonlinear
rather than linear. These studies can be classied into two categories. The rst is about econometric modeling and testing for
nonlinearity. For example, Kim and Nelson (2006) and Boivin (2006) adopted time-varying parameter models to examine the U.S.
data during the latest 50 years, and found that the reaction of federal fund rate to ination is obviously unstable, and what's more,
the U.S. monetary policy began to step into a more active state since 1980s. Rabanal (2004) estimated a two-state Taylor rule to
explain the behavior of the Federal Reserve over business cycle phases, and found that during expansions the Fed follows a rule
that can be characterized as ination targeting, whereas the Fed targets output growth during recessions. Bruggemann and Riedel
(2008) used logistic smooth transition regression (LSTR) models with time varying parameters to estimate the U.K. monetary
policy reaction function, and found that in times of recessions the Bank of England put more weight on the output gap and less on
ination, and a reverse pattern is observed in non-recession periods. The second is about theoretical analysis and explanation in
terms of three different ways, e.g. the aggregate supply curve being nonlinear, the central bank's preferences being asymmetric
around the targeted rate, and the central banker facing uncertainty about the model describing the economy. For example, Dolado,
Maria-Dolores, and Naveria (2005) relaxed the linear assumption of Phillips curve, and assumed that ination is a concave
function of output gap, which means a nonlinear Phillips curve. Nobay and Peel (2003), Ruge-Murcia (2003), Surico (2007a, b)
1
set
nonlinear welfare loss functions of central bank, which can represent central bank's asymmetric preference to positive or negative
deviations of ination and output gap from its target. Tillmann (2010) discussed the optimal monetary policy under parameter
uncertainty, and showed that if the central bank is uncertain about the slope of Phillips curve and applied monetary policy
according to a minmax strategy, then the response of interest rate to ination becomes stronger when ination is away from its
target. In addition, Davig and Leeper (2007) generalized the Taylor principle to an environment in which reaction coefcients in
the monetary policy rule evolve according to a Markov process.
Fromthe practical situation of China's economic development, there are still a lot of controversies about whether the interest
rate or the money supply is more appropriate as an intermediate goal for monetary policy. If Taylor rule may track China's
monetary policy well, then taking the interest rate as an intermediate target will offer the central bank a wider selection of policy
targets. So far, the Chinese scholars have done a lot of explorations and discussions. Xia and Liao (2001), after having examined
the conduct of China's monetary policy in the past, have concluded that the quantity control of money supply is not suited any
more for the Chinese economy. Xie and Luo (2002) rstly employed the historical analysis and reaction function method to
conduct an empirical analysis of China's monetary policy in the framework of Taylor rule and draw the conclusion that this rule
can accurately measure the operation level of China's monetary policy. Zhao and Gao (2004) constructed a more suitable interest
rate rule in China by allowing the exchange rate to inuence the long-run ination target. Bian (2006) used both general method
of moments (GMM) and cointegration test method to examine the applicability of Taylor rule in China. Considering standard
Taylor rule and the optimal monetary policy raised by Clarida, Gali, and Gertler (2002) in a comprehensive way, Wang and Zou
(2006) proposed an extensive Taylor rule in the opening-up economic condition and made empirical test on China's monetary
policy. In short, these studies have largely promoted and enriched the understanding of monetary policy rules for China.
However, the estimated results of policy reaction functions showed unstable behavior of China's monetary policy, which is not
consistent with the traditional Taylor rule. A more likely explanation is that these studies specied the policy reaction functions
improperly.
Recently, a few studies attempt to employ nonlinear methods to investigate China's monetary policy reaction function. Zhang
and Zhang (2008) applied a threshold regression model and found that all reaction coefcients in the regime of high money
growth are larger than those in the regime of low money growth. Ouyang and Wang (2009), based on the situation of asymmetric
reaction of monetary policy to ination and real GDP, constructed a nonlinear monetary policy reaction function using economic
growth rate and ination rate as threshold variables. The results showed that the reaction of monetary policy to ination and GDP
gap is nonlinear and asymmetric. Zheng and Liu (2010) developed a regime switching Taylor rule with time-varying ination
target and showed that China's monetary policy has signicant regime switching feature, that is, different regimes reect different
monetary policy reactions. However, possible two drawbacks of their study are worthy of consideration: the rst is that the
estimated model is not based on the forward-looking Taylor rule, so it might be inconsistent with the People's Bank of China
(PBC)'s actions; the second is that they do not deal with the endogeneity problem in the regime switching model which may lead
to inconsistent parameter estimates for Taylor rule.
In this study, we consider a regime-switching forward-looking Taylor rule that nests the simple rule with regime switching in
Zheng and Liu (2010) as a special case. For this regime switching model, the traditional generalized method of moments (GMM) is
1
Surico (2007b) showed that the rst six years of ECB monetary policy can be characterized by a nonlinear, state-dependent policy rule. The empirical analysis
on synthetic euro-area data suggests that the objective of price stability is symmetric, whereas the objectives of real activity and interest-rate stabilizations are
not. Output contractions imply larger policy responses than output expansions of the same size, while no asymmetric reaction for ination.
48 T. ZHENG et al. / China Economic Review 23 (2012) 4759
no longer suitable for estimating this type of nonlinear reaction function, since the forward-looking rule has a dynamic structure
with endogenous regressors. Therefore, we realize its appropriate estimation of this nonlinear rule via the two-step MLE
procedure due to Kim and Nelson (2006), Kim (2009) and Zheng and Wang (2010) to deal with the potential endogeneity
problem.
According to formal econometric methods, we would select two-regime Markov-switching forward-looking Taylor rule to
investigate China's monetary policy reaction function from 1992 to 2010. As far as we know, the forward-looking rule is more
reasonable than the simple rule due to at least the following three reasons. Firstly, it's helpful to increase the degree of
transparency and accountability and thus maintain stable economic growth (Svensson, 1999) and it's also useful to stabilize public
expectation and improve social welfare (Liu, 2004). Secondly, the forward-looking rule is more consistent with the PBC's actions.
The fact that the central government determines the coming year's monetary policy at the end of each year illuminates the
important role of expected component in the PBC's monetary policy decision. Thirdly, the state council executive meeting on 21st
Oct. 2009 formally proposed management of ination expectations, which conrms that the PBC targeted on expected ination
rather than past or actual ination.
The remainder of the paper is organized as follows. Section 2 briey reviews traditional and forward-looking Taylor rules.
Section 3 introduces a regime-switching forward-looking Taylor rule and presents its estimation procedure. Section 4 describes
the data selection and processing. In Section 5, we implement formal econometric testing and model selection, and report the
empirical results for China's monetary policy. Finally, the last section draws some conclusions.
2. Linear Taylor rules
2.1. Traditional Taylor rule
Taylor rule, proposed by Taylor (1993) for U.S. real data during the period from 1987 to 1992, is one of the simplest monetary
policy rules in common use. Given the ination target and potential output, it provides an adjustment criteria of short-term
interest rate to the changes of ination and real output.
Taylor (1993) assumes that the central bank follows the following general behavior criteria:
i

t
= r +

+
t

_ _
+ y
t
y

t
_ _
; 1
where i
t
* is the desired value of the short-term nominal interest rate. It is assumed to rise if ination rate () rises above its target
(*) or if real output (y) increases above its trend (y*). Therefore, in Eq. (1), the parameter indicates the sensitivity of interest
rate policy to the deviation of ination rate from its target, and the parameter indicates the sensitivity of interest rate policy to
the output gap. However, once the economy reaches the long-run equilibrium, the deviation of ination rate from its target value
is zero and the output gap is zero. Consequently, the desired nominal interest rate (i*) is simply the sum of the equilibrium real
interest rate r and the target value of ination (*), reducing to the standard Fisher Equation.
An important empirical question is related to the estimated weight on ination. Since it is the real interest rate that actually
drives private decisions, the size of needs to assure that the nominal interest rate is raised enough to actually increase the real
interest rate. This so-called Taylor principle implies that the coefcient on the ination gap () should exceed 1.
2
A coefcient
larger than one on ination means that the central bank increases the real interest rate in response to higher ination, which
exerts a stabilizing effect on ination; on the other hand, b1 indicates an accommodative behavior of interest rate to ination,
which may generate self-fullling bursts of ination and output.
3
In addition, the coefcient on the output gap () should be
positive. A positive coefcient on output gap means that if real output is belowits potential level, a decrease in the interest rate will
have a stabilizing inuence on the economy.
In his study on U.S. data, Taylor (1993) suggested that both the equilibriumreal interest rate and the target ination rate be 2%.
He also assumed that the weight of ination gap and output gap relative to Federal Funds rate might be 1.5 and 0.5, respectively,
i.e. =1.5,=0.5.
2.2. Forward-looking Taylor rule
In the traditional Taylor rule, the reaction of interest rate to ination and output gap is contemporary or backward-looking. For
example, Taylor (1993) considered the deviation of ination from target over the last four quarters. However, in practice, central
banks do not tend to take the past or actual ination as the target but the expected ination. Therefore, Clarida et al. (1998)
suggested introducing the expectation to construct a forward-looking version of Taylor rule, which allows the central bank to
consider a broad array of information to form beliefs about the future condition of the economy. According to Clarida et al. (1998,
2000), the central bank's desired target interest rate (i
t

) depends not only on the deviation of k periods ahead expected ination


2
See, e.g. Taylor (1999) and Clarida et al. (1998).
3
Benassy (2006) characterized optimal interest rate rules in the framework of a dynamic stochastic general equilibrium model, and found that the elasticity of
response for interest rate depends on the degree of price rigidity, the autocorrelation of the underlying shocks, or which measure of ination is used. It is also
suggested that in general the optimal elasticity of the interest rate with respect to ination needs not be great than one.
49 T. ZHENG et al. / China Economic Review 23 (2012) 4759
(in annual rates) from its target value but also the p periods ahead expected output gap. Then we have the following forward-
looking Taylor rule
4
:
i

t
= r +

+ E
t

t;k
_ _

_ _
+ E
t
y
t;p
y

t;p
_ _
; 2
where E
t
is the conditional expectation operator, conditional on information available to the monetary authority at time t, E
t
(
t, k
)
is the k periods ahead forecast for ination at time t, and E
t
(y
t, p
y
t, p
* ) is the p periods ahead forecast for output gap at
time t.
Moreover, we also assume that the central bank adjusts interest rates in a cautious way through smoothing in the form of
partial adjustment as follows
5
:
i
t
= 1 i

t
+ i
t1
+ m
t
3
where (0, 1) is the smoothing parameter, m
t
is a random disturbance term caused by the central bank's control of interest rate.
According to this partial adjustment behavior, the central bank at each period adjusts its instrument in order to eliminate only a
fraction (1) of the gap between its current target level and some linear combination of its past values. Therefore, the parameter
can be considered as an index which captures the degree of interest rate smoothing (Clarida et al., 2000).
Then, let

y
t;p
= y
t;p
y

t;p
denote the output gap, c = r 1

, and substitute Taylor rule Eq. (2) into Eq. (3). Rearranging,
we can get the following simplied forward-looking monetary policy function:
i
t
= 1 c +
t;k
+

y
t;p
_ _
+ i
t1
+ e
t
4
where e
t
= 1
t;k
E
t

t;k
_ _ _
+

y
t;p
E
t

y
t;p
_ _ _ _ _ _
+ m
t
.
6
Eq. (4) is regarded as the forward-looking monetary reaction
function considering interest rate smoothing. Kim and Nelson (2006) showed that there may be some dependence between the
error term (e
t
) and two explanatory variables (
t, k
and

y
t;p
), which produces the problem of endogeneity.
In the existing literature, Eq. (4) is usually estimated by the generalized method of moments (GMM). According to Clarida et al.
(1998, 2000), this estimation method is well suited for the econometric analysis of interest rate rules when the regressions are
made on that are not known by the central bank at the decision-making moment. To implement this estimation procedure, the
following orthogonal condition is imposed on:
E
t
i
t
1 c +
t;k
+

y
t;p
_ _
i
t1
j z
t
_ _
= 0; 5
where z
t
is a vector of instrument variables that are orthogonal to e
t
. These variables can be a set of lagged variables and additional
exogenous variables that help to predict ination and output gap. In addition, in order to account for the possible
heteroscedasticity and serial correlation in e
t
, an optimal weighting matrix is often used in estimation. Therefore, the GMM
estimator is robust to heteroscedasticity and serial correlation.
In Kim and Nelson's (2006) study, they applied Kim's (2006) time-varying parameter model with endogenous regressors to
model the forward-looking Taylor rule. Their work made two important progresses. First, it can be extended to deal with
nonlinearities, which allow all parameters in the model to be time-varying. Second, it can deal with heteroscedasticity in the
disturbance terms for the monetary policy rule, which allows us to consider the degree of uncertainty associated with economic
conditions. In order to obtain the consistent estimation in the time-varying parameter model, they suggested a two-step MLE
procedure to deal with endogeneity in regressors.
3. Regime-switching Taylor rule and its estimation
In this section, the linear forward-looking Taylor rule is generalized to the regime switching form. We rst introduce the
regime switching forward-looking Taylor rule, and then set up its econometric model. After that, we apply the two-step MLE
procedure suggested by Kim and Nelson (2006) and Kim (2004, 2006, 2009) and Zheng and Wang (2010) to estimate this
nonlinear model.
4
Although empirically motivated, the forward-looking Taylor rule can also be derived theoretically. Assuming that the evolution of economy described by the
New-Keynesian model is linear, that is the linear IS curve and the linear Phillips curve, and the central bank uses a policy rule to implement monetary policy, then
the central bank can get the rst-order necessary conditions for the policy reaction function (i.e. the forward-looking Taylor rule) by minimizing the inter-
temporal loss function.
5
Traditional explanations for the interest rate smoothing hypothesis include: fear of disrupting capital markets, loss of credibility from a sudden large policy
reversal, the need for consensus building for a policy change, etc. In addition, interest rate smoothing may be thought of as a learning device by the central bank,
which may not have a full knowledge of economy due to imperfect information.
6
When expected value equals actual value, i.e. expected ination equals to actual ination and expected output gap equals to actual output gap, this forward-
looking Taylor rule degenerates into the contemporary Taylor rule.
50 T. ZHENG et al. / China Economic Review 23 (2012) 4759
3.1. Model specication
For the linear Taylor rule, policy reaction coefcients and interest rate smoothing parameter are generally assumed to be
constant. But in the long run, these parameters may be time-varying or state dependent. As mentioned earlier, we consider the
following extended forward-looking Taylor rule with regime switching nonlinearity:
i
t
= 1
S
t
_ _
c
S
t
+
S
t

t;k
+
S
t

y
t;p
_ _
+
S
t
i
t1
+ e
t
6
where e
t
= 1
St
_ _

St

t;k
E
t

t;k
_ _ _
+
St

y
t;p
E
t

y
t;p
_ _ _ _ _ _
+ m
t
is the disturbance term. S
t
is a discrete random variable,
whose realization can be interpreted as the state of economy at period t. In this study, we assume that the state variable S
t
can be 1,
2, N, and it follows a rst-order Markov chain with the constant transition probability matrix:
P =
p
11
p
21
p
N1
p
12
p
22
p
N2

p
1N
p
2N
p
NN
_

_
_

_
7
where p
ij
=Pr(S
t
=j|S
t 1
=i) is the transition probability of moving fromthe state S
t 1
=i at time t 1 to the state S
t
=j at time t,
i, j =1, 2, , N.
Eq. (6) relaxes the linearity assumption of the traditional Taylor rule, allowing all parameters to be possibly state dependent,
i.e. the parameters c, , and can switch across states. By taking the policy reaction coefcients as the regime switching form, we
can capture possible regime shifts in monetary policy operations.
In addition, as argued by Sims (2001) and Sims and Zha (2006), we note the importance of time-varying variance of the shocks
in the monetary policy rule. Thus, the disturbance term e
t
is assumed to follow a Gaussian distribution with state-dependent
variance, i.e.
7
e
t
N 0;
2
S
t
_ _
; 8
and without loss of generality, we assume that
1
2
N
2
2
NN
N
2
.
In order to estimate above Eqs. (6)(9), we need several instrumental variables (IV). Similar to Kim and Nelson (2006), we
assume that the relationships between the endogenous regressors (
t, k
and

y
t;p
) in Eq. (6) and the vector of IV z
t
are given
by:

t;k
= z

1t
+ v
1t
; v
1t
N 0;
2
v1
_ _
; 9

y
t;p
= z

2t
+ v
2t
; v
2t
N 0;
2
v2
_ _
; 10
with

it
=
it1
+ u
it
; u
it
iidN 0;
u;i
_ _
; i = 1; 2; 11
where the error terms v
1t
and v
2t
are assumed to be correlated with the correlation coefcient , i.e. corr(v
1t
, v
2t
)=.
As specied above, the relationships between the endogenous regressors (
t, k
and

y
t;p
) and the vector of IVz
t
in Eq. (6) could be
time-varying. So, this time-varying parameter model could catch some gradual changes over time.
3.2. Two step estimation procedure
Nowlet us take in account the estimation of the regime switching model given by Eqs. (6)(8) mentioned above. Obviously, as
an unobservable discrete variable is added, it is not possible to apply the GMM estimation procedure to estimate this regime
switching model. Therefore, this study estimates this nonlinear model with endogenous repressors based on the two-step MLE
procedure suggested by Kim and Nelson (2006), Kim (2009) and Zheng and Wang (2010).
8
7
An alternative possible setting for the time varying variance is using the GARCH (generalized autoregressive conditional heteroscedasticity) models. See for
example Kim and Nelson (2006) and Kim, Kishor, and Nelson (2006).
8
The two-step estimation procedure has at least two advantages. One is that it provides us a consistent estimator rather than an approximate estimator due to
Kim's (1994) approximation (Zheng & Wang, 2010). Another important advantage is that as the number of states for the latent Markov-switching variables and
endogenous regressors, the joint estimation procedure is infeasible due to the curse of dimensionality (Kim, 2009). In this case, we have no choice but to
employ the two-step procedure introduced in this paper, at the cost of a loss of asymptotic efciency.
51 T. ZHENG et al. / China Economic Review 23 (2012) 4759
For the model given by Eqs. (9)(11), we can represent it as the following state space form:

t;k

y
t;p
_ _
= I
2
z

t
+ v
t
= I
2
z

1 = 2
v

t
v

t
_ _
;
1 = 2
v
=

1
0

2

2
_ _
;
_ _
12

t
v

t
_ _
=
I
k
0
0 0
_ _


t1
v

t1
_ _
+
u
t
v

t
_ _
;
u
t
v

t
_ _

N
0
0
_ _
;

u
0
0 I
2
_ _ _ _
; 13
where the time varying parameter vector
t
=
1t

2t

, the mutually independent error vector v

t
= v

1t
v

2t
_

are iid
standardized normal variables. It is obvious that the model given by Eqs. (12)(13) can be estimated with the maximumlikelihood
estimation procedure by running the conventional Kalman lter. Furthermore, smoothed estimates of

t j T
and its corresponding
variance

P
t j T
can also be obtained from the conventional Kalman lter. We denote the last K1 block of

t j T
as v
t|T

, and the last


KK block of

P
t j T
as P
t|T

=E[(v
t

v
t|T

)(v
t

v
t|T

)], which will be used in the second step.


Since there exists correlated structure between the shock terme
t
and two regressors
t, k
and

y
t;p
, we then assume the following
covariance structure between v

t
= v

1t
v

2t
_

and e
t
:
v

t
e
t
_ _

N
0
0
_ _
;
I
2

S
t

S
t

S
t

S
t

2
S
t
_ _ _ _
14
where
St
=
1
S
t

2
S
t
is a state-dependent 21 vector of correlation coefcients. As shown in Kim(2004, 2009), the Cholesky
decomposition of the covariance matrix of v

t
e
t
_

in Eq. (14) results in the following representation:


v

t
e
t
_ _
=
I
2
0

S
t

S
t

1
S
t

S
t

S
t
_


t

t
_ _
;

t

t
_ _

i:i:d:N
0
0
_ _
;
I
2
0
0 1
_ _ _ _
_
15
where I
2
is a 22 identity matrix.
Then, e
t
can be rewritten as a linear form of v
1t
* and v
2t
* , i.e.
e
t
=
S
t
v

t
+
t
;
t
N 0;
2
;S
t
_ _
; 16
where
S
t
=
S
t

S
t
,
, S
t
2
=(1
S
t

S
t
)
S
t
2
, and the disturbance term
t
is uncorrelated with either v
1t
* or v
2t
* . Thus, substituting
Eq. (16) into Eq. (6), we get the transformed model as follows:
i
t
= 1
S
t
_ _
c
S
t
+
S
t

t;1
+
S
t

y
t;1
_ _
+
S
t
i
t1
+

S
t
v

t
+
t
: 6
As the disturbance term
t
is uncorrelated with
t, 1
,

y
t;1
, v
1t
* and v
2t
* , the following two-step MLE would be valid:
Step 1: Estimate Eqs. (9)(11) via the MLE procedure based on the conventional Kalman lter, and obtain standardized errors

1;t j T
and

v

2;t j T
, and their corresponding covariance matrix

P

t j T
= E v

t
v

t j T
_ _
v

t
v

t j T
_ _

_ _
.
Step 2: Using the maximum likelihood method via the Hamilton's (1989) lter to estimate the following equation:
i
t
= 1
S
t
_ _
c
S
t
+
S
t

t;1
+
S
t

y
t;1
_ _
+
S
t
i
t1
+
S
t

t j T
+

t
;

t
i:i:d:N 0;
2
t
_ _
; 6
where

t
=
t
+
St
v

t j T
_ _
and
2
t
= 1
St
I
2

t j T
_ _

St

2
St
_
which is suggested by Zheng and Wang (2010). Given

v

t
and
the corresponding covariance matrix

P

t j T
, the model parameters can be estimated by the maximum likelihood procedure, and
then the unobservable discrete variable S
t
can be exactly inferred given estimated parameters to get the ltered and smoothed
probabilities.
Remark 1. In the maximum likelihood estimation of Eq. (6) based on the derived Hamilton lter, the parameter estimates are
consistent, and the standard errors have been corrected by introducing the term

P

t j T
into the variance of
t
* for Pagan's (1984)
generated regressors' problem. For further details, the readers are referred to a series of papers by Kim (2004, 2006, 2009), Kim
and Nelson (2006), Kim and Kim (2007) and Zheng and Wang (2010).
Remark 2. The aforementioned two step estimation procedure can also be used to estimate the model of linear Taylor rule. If one
takes account of one regime or the state variable only takes one number, then the regime switching Taylor rule would reduce to a
linear rule. In the case of the two step estimation for the linear rule, the sole difference from that of the regime switching rule is
that all parameters are not state dependent, leading to a very simple MLE procedure in the second step.
52 T. ZHENG et al. / China Economic Review 23 (2012) 4759
4. Data
In this study, the data series we employed are quarterly data for China covering the period 1992:Q1-2010:Q3 with 75
observations in total. The detail of data selection, processing, and description is given as follows:
(1) Short-term nominal interest rate. According to most of Chinese scholar's studies such as Xie and Luo (2002), Lu and Zhong
(2003), Zhao and Gao (2004), and recently Zheng and Liu (2010), we choose the China inter-bank offered rate (CHIBOR) as
a proxy variable of interest rate. This index meets with the requirement of market interest rate, and can quickly reect the
demand and supply in money market. The data of quarterly interest rate during the period from 1992 to 1995 is the inter-
bank offered rate in Shanghai Financial Center. Although the lending behavior of China's nancial institutions is disordered
before and after 1993, the inter-bank offered rate in Shanghai can still reect the situation of nationwide inter-bank offered
market before networking in 1996 (Xie & Luo, 2002). For the interest rate data during the period 19962010, we select 7-
day China inter-bank offered rate,
9
which can essentially reect the variability of recent market position. We get quarterly
CHIBOR by quarterly averaging of monthly data, which are published by the PBC's (People's Bank of China) Quarterly
Statistical Bulletin and the PBC's website (http://www.pbc.gov.cn).
10
The nal data of quarterly short-terminterest rate are
depicted in Fig. 1(a).
(2) Ination rate. Taking into account the availability and reliability of consumer price index (CPI) and GDP deator, we choose
the CPI as a measure of ination rate. In our study, the annualized quarterly ination rate is calculated according to the data
of monthly year-on-year growth rate of CPI, which is obtained from CEI database (available from: http://db.cei.gov.cn) and
China Monthly Economic Indicators. Since the ofcial CPI is monthly year-on-year growth rate, three term moving average
1992 94 96 98 2000 2 4 6 8 10
2.5
5.0
7.5
10.0
12.5
a) Interest rate
1992 94 96 98 2000 2 4 6 8 10
0
5
10
15
20
25
b) Inflation
1992 94 96 98 2000 2 4 6 8 10
-10
-5
0
5
10
c) Output gap
1992 94 96 98 2000 2 4 6 8 10
10
20
30
40
50
d) M1 growth
Fig. 1. Data series during the period 1992Q1 to 2010Q3.
9
Due to the limited data sources, the weighting interest rates during the period 19921995 are calculated by all term interest rate weighting of Shanghai
Financial Center, whereas for the period from 1996 to 2009 the interest rate is seven day inter-bank offered rate. Although the term in these two periods does not
match, it has little effect on modeling in that there is not large in interest differential of different terms for Shanghai Financing Center. (Xie & Luo, 2002).
10
Based on the data of the monthly inter-bank offered rate and its corresponding trading volume, we can calculate the quarterly weighted average inter-bank
offered rate as follows:
i = i
1
f
1
f
+ i
2
f
2
f
+ + i
n
f
n
f
=
if
f
;
where i
k
is the monthly interest rate, f
k
is its corresponding trading volume.
53 T. ZHENG et al. / China Economic Review 23 (2012) 4759
is adopted to get quarterly year-on-year growth rate of CPI, and then the quarterly data of annual ination rate,
t
, is
calculated as follows:
t
=(quarterly CPI 1)100%. The quarterly ination rate based on CPI is shown in Fig. 1(b).
(3) Output gap. To get the real output gap, we should rstly measure potential output based on real output. The gross domestic
product (GDP) is selected as a measure of total output. Ofcial statistical data provides quarterly GDP at current prices and
cumulative GDP growth rate over the same period last year since 1992. The data is available from CEI database (http://db.
cei.gov.cn) and China Monthly Economic Indicators. In order to get real GDP, denoted by Y
t
, we rst re-calculate real value at
comparable price (setting year 2000=100) based on cumulative GDP growth rate, and then compute the seasonally
adjusted GDP series via X-12 seasonal adjustment method executed by the software Eviews. After applying the Hodrick
Prescott lter (HP lter) to obtain potential output, denoted by Y
t
*, the annualized output gap is computed as
percentualized log-deviation of real output with respect to potential output, i.e. y
t
400log(Y
t
/Y
t
*). Fig. 1(c) draws the nal data
of real output gap.
In addition, to estimate the forward-looking monetary policy reaction function, we need to construct instrumental variables
(IV) to calculate standardized prediction errors {v
1t

} and {v
2t

}. The IV include four lags of each of the following variables: the short-
term inter-bank offered rate, ination rate, output gap, and the M1 growth rate which is shown in Fig. 1(d).
11
5. Empirical results
In this section, we rst consider the selection of the model, then estimate the regime switching forward-looking Taylor rule,
and at last analyze the asymmetric behavior of China's monetary policy.
5.1. Model selection
5.1.1. Determination of k and p
The rst task of model selection is to determine the appropriate numbers of k and pfor the linear forward-looking rule. We
estimate the linear model with different numbers of k and p (i.e. taking the values 0, 1, 2, 3 or 4) using the two step estimation
procedure, and then choose the model which minimizes the information criteria such as Akaike's information criterion (AIC) and
Bayesian information criteria (BIC). The smaller is the AIC or the BIC, the better the model is.
Table 1 presents the AICs and BICs of the linear Taylor rule with different numbers of k and p. According to the results, both the
AIC and BIC values arrive at the minimum value when k=p=2. Therefore, in the following context, we consider this appropriate
setting k=p=2 and x them.
The parameter estimates of linear forward-looking Taylor rule are reported in Table 2. As shown in the table, the interest rate
smoothing parameter is very signicant, indicating that the interest rate adjustment mechanismstrongly depends on the previous
interest rate. For the response of short-term interest rate to ination, the policy parameter is signicant but less than 1,
indicating that the interest rate policy rule is unstable; moreover, for the response of short-terminterest rate to the output gap, the
policy parameter () is insignicant. The above results are consistent with a majority of Chinese scholars' researches, such as Xie
and Luo (2002) and Bian (2006). In addition, the estimation results of a linear model without considering endogenous regressors
are also compared with the two step estimation results. The results showthat the correlation coefcient with respect to output gap
is signicant, showing its importance of considering the problem of endogeneity.
5.1.2. Determination of the number of regimes
For the specication of the Markov-switching model, it is important to know the number of states which is often not known a
priori. Although Akaike's information criterion (AIC) has been used in various model selection contexts, Smith, Naik, and Tsai
(2006) found that in Markov-switching models it misleads the users into selecting too many states. Using the KullbackLeibler
Table 1
AICs and BICs under different choice of k and p.
AIC BIC
p=0 p=1 p=2 p=3 p=4 p=0 p=1 p=2 p=3 p=4
k=0 111.86 102.05 111.07 106.89 105.76 127.70 117.79 126.71 122.42 121.19
k=1 106.70 108.19 105.10 103.04 103.00 122.44 123.93 120.73 118.58 118.43
k=2 100.09 99.56 90.16 98.62 99.16 115.73 115.20 105.80 114.15 114.59
k=3 109.01 100.32 99.41 99.90 98.93 112.06 115.86 114.95 115.44 114.37
k=4 105.76 103.00 99.16 98.93 95.41 121.19 118.43 114.59 114.37 110.84
Note: The gures are the corresponding Akaike information criterions (AIC) and Bayesian information criterions (BIC) computed by the linear model based on
different values of k and p.
11
We do not adopt the M2 growth rate, mainly because the ofcial provided M2 growth rate since 1996.
54 T. ZHENG et al. / China Economic Review 23 (2012) 4759
divergence between the true and candidate models, Smith et al. (2006) derive the following formal procedure called Markov
switching criterion (MSC) to select the number of states and variables simultaneously:
MSC = 2log L +
N
i =1
T
i
T
i
+
i
K

i
T
i

i
K2
17
where N is the number of states and K is the number of the regressors in the model. Also, T
i
is the sumof smoothed probabilities of
being in the regime i.
i =1
N
T
i
=T, where T is the number of the effective observations used in estimation. The second term in Eq.
(17) is a penalty for the complexity of a model. Smith et al. (2006) recommend setting
i
=1 and
i
=N (also
i
=1 or
i
=N
2
is
another possible choice but they show that MSC
=1
and MSC
=N
2 perform worse than MSC
=N
when both the sample size is
small and the signal is weak).
We estimate one-state, two-state and three-state Markov-switching models, respectively, and compute two estimates of KL
divergence: AIC and MSC
=N
. Based on the minimum MSC
=N
value, we would select a model with N

=2 while the minimum


value of AIC yields N

=3. This nding is consistent with the simulation evidence given by Smith et al. (2006) which reveals AIC's
tendency to select more states than necessary. Therefore, this paper takes N=2 as the number of Markov-switching states.
5.2. Results of the regime switching model
Table 3 reports the estimation results of the regime-switching forward-looking Taylor rule with two states. As shown in the
table, all parameter estimates are signicantly different from zero under 5% signicant level, except for c
1
,
1
and
12
(the
correlation coefcient
1
in the second regime or Regime 2). It seems that the regime switching model can well describe the
dynamic behavior of interest rate policy. Moreover, for the regime switching model, its log likelihood value increases by 39.89
relative to the linear model, showing better performance in modeling the short term rate. In particular, the results also show
important implication of considering the problem of endogeneity. As we can see from the table, the correlation coefcients
2
in
both regimes are signicant and very high, while the correlation coefcient
1
in the second regime is not signicant, but
signicant in the rst regime. Therefore according to the two-step MLE procedure, we can also capture the existence of signicant
correlation in endogenous structure using the regime switching model.
Table 2
Parameter estimates of linear Taylor rule.
Parameter Estimation without endogeneity Two-step estimation
Estimate Std error p-value Estimate Std error p-value
Interest rate smoothing 0.9390 (0.0160) [0.0000] 0.9178 (0.0165) [0.0000]
Constant c 0.5099 (1.6482) [0.7581] 0.3913 (1.0114) [0.7002]
Ination reaction 0.9710 (0.2356) [0.0001] 0.8537 (0.1434) [0.0000]
Output gap reaction 0.1405 (0.2349) [0.5518] 0.2743 (0.1807) [0.1340]
Correlation coefcient
1
0.1743 (0.1338) [0.1975]
Correlation coefcient
2
0.5381 (0.1186) [0.0000]
Standard deviation 0.4603 (0.0395) [0.0000] 0.4684 (0.0407) [0.0000]
Log-likelihood 43.7275 38.0813
Note: Figures in parentheses and brackets are corresponding standard errors and p-values of parameter estimates.
Table 3
Parameter estimates of regime-switching Taylor rule.
Parameter Regime 1 Regime 2
Estimate Std error p-value Estimate Std error p-value
Interest rate smoothing 0.8894 (0.0223) [0.0000] 0.9474 (0.0082) [0.0000]
Constant c 1.6846 (1.6420) [0.3094] 2.3796 (0.5075) [0.0000]
Ination reaction 1.3312 (0.2362) [0.0000] 0.3683 (0.0687) [0.0000]
Output gap reaction 0.2499 (0.2049) [0.2280] 0.7333 (0.1439) [0.0000]
Correlation coefcient
1
0.4610 (0.1427) [0.0021] 0.1138 (0.0938) [0.2307]
Correlation coefcient
2
0.7364 (0.1268) [0.0000] 0.9460 (0.0587) [0.0000]
Standard deviation 0.4398 (0.0622) [0.0000] 0.1891 (0.0240) [0.0000]
Transition probability matrix
Pr(S
t
=1|S
t 1
) 0.7512 (0.1124) [0.0000] 0.1449 (0.0675) [0.0363]
Pr(S
t
=2|S
t 1
) 0.2488 (0.1124) [0.0312] 0.8551 (0.0675) [0.0000]
Log-likelihood 1.8082
MSC
=N
161.67
Note: Figures in parentheses and brackets are corresponding standard errors and p-values of parameter estimates.
55 T. ZHENG et al. / China Economic Review 23 (2012) 4759
However, two important issues arise, i.e. whether the interest rate policy takes on regime switching nonlinearity, and whether
interest rate smoothing parameter, policy reaction coefcients and standard deviation are asymmetric in different regimes. In
order to check these problems, we carry out the following two formal tests, respectively.
5.2.1. Testing for linearity
Including regime switching nonlinearity appears to represent an improvement over the linear model, comparing Table 2 and
Table 3. Testing the restriction of the linear model, i.e. that all parameters are not state dependent, the likelihood ratio (LR) test
statistic is 79.78. This test statistic, however, is nonstandard.
12
In order to establish the statistical signicance of this result, a
parametric bootstrap test was performed in terms of Di Sanzo (2009).
13
The bootstrap algorithmruns as follows: The data were simulated under the null of no regime switching nonlinearity, i.e. using
the parameter estimates of the linear model fromTable 2. The model was re-estimated for each sample under both the null and the
alternative to obtain a likelihood ratio test statistic. The bootstrapped p-value, based on 10,000 bootstrap samples, is 0.0001, which
is the fraction of simulated LR values that are greater than the observed LR value. This suggests that regime switching nonlinearity
is indeed appropriate for explaining the movements in China's monetary policy reaction.
5.2.2. Testing parameter stability
After estimating the regime switching model, we proceed to test the possible asymmetries in the policy rule. By imposing
restrictions on the parameters being not state dependent, we test the following null hypotheses of parameter stability:
1
=
2
,
where could be the interest rate smoothing parameter, policy reaction coefcients, correlation coefcients or standard deviation.
The statistical signicance of the restrictions imposed on the model is assessed by the usual Wald test. Under null hypothesis, the
Wald test has a chi-square distribution with 1 degree of freedom.
14
Table 4 reports the Wald test results of parameter stability across regimes. For the interest rate smoothing parameter (), policy
reaction coefcients ( and ) and standard deviation (), each Wald test for parameter stability shows that the null hypothesis
can be strongly rejected under 5% signicant level, indicating that these parameters are statistically different between two
regimes. For the correlation coefcients (
1
and
2
) capturing endogenous structure, the Wald test results show that the
correlation coefcient of interest rate to the standardized prediction error of expected ination (v
1

) is signicantly state
dependent, while the correlation coefcient of interest rate to the standardized prediction error of expected output gap (v
2

) is not
state dependent.
5.2.3. The policy reaction function
In this part, let us turn to evaluate the nonlinear policy reaction function according to the parameter estimates according to
Table 3 and Table 4. From the two parameters and , the result reveals that the interest rate smoothing mechanism and the
interest rate shock are variant, depending on the state of economy; furthermore, the greater the degree of interest rate smoothing
is, the smaller the variance of the interest rate shock is, and vice versa. Similarly, the standard deviation also depends on the
discrete state, implying that the variance of the interest rate shock is time-varying, but here it is represented in a regime switching
form. From policy reaction coefcients and , the reactions of interest rate policy to ination and output gap do not follow the
traditional Taylor rule. This is because in the second regime, the ination coefcient is less than 1 (b1) and the output gap
coefcient is larger than 0 (N0); however, in the rst regime, the ination coefcient is larger than 1 and the output gap
coefcient is not statistically different from zero.
Consequently, we can draw a conclusion that the monetary policy reaction function that relates the nominal interest rate to
ination and output uctuation is nonlinear and asymmetric. It is revealed in this paper that the PBC targets ination in the rst
Table 4
Wald test results.
W
1

1
=

2
_ _
W
2

c
1
=

c
2
_ _
W
3

1
=

2
_ _
W
4

1
=

2
_ _
W
5

11
=

12
_ _
W
6

21
=

22
_ _
W
7

1
=

2
_ _
6.5435 5.9510 15.562 4.2095 11.974 2.3825 14.405
[0.0105] [0.0147] [0.0001] [0.0402] [0.0005] [0.1227] [0.0001]
Note: The Wald statistic, W, is used to test the parametric stability of the regime switching model, i.e. whether a parameter in rst regime (
1
) is equal to another
parameter in the second regime (
2
), where the null hypothesis is
1
=
2
, the alternative hypothesis is
1

2
. Then Wald-test statistic is given by
W
1
=
2
=

1

2

2
Var
1
+ Var
2
2Cov
1
;
2

2
1 ;
where Var() is the variance of , Cov(
1
,
2
) is the covariance of
1
and
2
,
2
(1) is the chi-square distribution with freedom of degree 1.
12
Testing for linearity in the context of Markov-switching models is complicated because standard regularity conditions for likelihood based inference are not
maintained under the null hypothesis due to the presence of nuisance parameters and singularity of the information matrix (Hansen, 1992). Thus, the asymptotic
distribution of the relevant likelihood ratio (LR) statistic does not possess the standard chi-square-distribution.
13
Although several alternative test methods such as Hansen (1992, 1996), Garcia (1998) and Carrasco, Liang, and Ploberger (2004) are available, this study
adopts the bootstrap-based LR test introduced by Di Sanzo (2009) for testing linearity in Markov-switching models mainly due to its relative simplicity. Di Sanzo
(2009) showed that a bootstrap of a likelihood ratio test statistic performs well for testing linearity in Markov-switching models.
14
In more detail, this test for the regime switching model is referred to Hamilton (1996).
56 T. ZHENG et al. / China Economic Review 23 (2012) 4759
regime while focuses on the output gap in the second regime. That is, in the rst regime, if the ination rate is 1% higher than its
target value, the real interest rate would raise by 0.33%, and the nominal interest rate would raise by 1.33%; if the ination rate is
1% lower than its target value, the real interest rate and the nominal interest rate would fall by 0.33% and 1.33%, respectively. But it
is not obvious for the output gap. In the second regime, if the real output is 1% higher than its potential level, the nominal interest
rate and real interest rate would raise by 0.73% simultaneously; and if the real output is 1% lower than its potential level, the
nominal interest rate and real interest rate would fall by 0.73% simultaneously. Under the output gap unchanged, if the ination
rate is 1% higher (lower) than its target value, the nominal interest rate would raise (fall) by 0.37%, while the real interest rate
would fall (raise) by 0.63%.
5.3. Regime switching behavior of China's monetary policy
To learn the regime switching characteristic of China's monetary policy, we further estimate the regime probabilities of the
state variable S
t
using parameter estimates of the regime switching model in Table 3. Fig. 2 depicts the time paths of the smoothed
probabilities in the rst regime, i.e. Pr(S
t
=1|
T
), t =1, , T, where
T
denotes all available information in the data sample.
According to the decision rule whether the probability is larger than 0.5, we also drawthe corresponding state by the shading area.
As shown in Fig. 2, China's monetary policy is obviously in the rst regime during the periods 1993:Q2-1993:Q3, 1995:Q1-
1995:Q3, 1997:Q4-1998:Q3, 1999Q3 and 2007:Q1-2009:Q3, while it is in the second regime during other periods. This result
implies that China's monetary policy experienced more periods in the second regime than those in the rst regime. This can also be
reected by the probabilities of self-maintenance, showing that the expected duration (D
i
=1/(1p
ii
)) of the second regime is
longer than that of the rst regime, about 4 and 7 quarters, respectively.
To show the relationship between economic situations and monetary policy intuitively, Fig. 3 demonstrates the time paths of
data series and the periods in the rst regime. Obviously, the occurrences of the rst regime mostly correspond to several major
adjustments in the benchmark of lending and deposit interest rates. It suggests that the central bank prefers to adjust interest rate
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Fig. 2. Smoothed probabilities in the rst regime.
Interest rate
Inflation
Output gap
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
-10
-5
0
5
10
15
20
25
Interest rate
Inflation
Output gap
Fig. 3. Data series and regime shifts.
57 T. ZHENG et al. / China Economic Review 23 (2012) 4759
to manage public ination expectations and thus curb deation or high ination effectively in the rst regime. Moreover, the
interest rate policy is always evident during the periods in the rst regime, taking on larger interest rate changes. For instance, in
1993 China began to appear the most serious ination since reform and opening up, the central bank increased the interest rate in
May and July in 1993 and January and July in 1995 respectively, and then controlled the ination effectively. When faced with
Asian nancial crisis and the clear signs of deation in the domestic economy in the end of 1997, the central bank decreased the
interest rate four times to curb the deepening deation trend from October 1997 to December 1998. Similarly, the central bank
increased interest rate six times continuously to prevent the evident ination in 2007, which was the result of excessive credit
growth and continuous price increase. However, as the once in a century international nancial crisis deeply inuenced China's
economy and caused deation, the central bank decreased the interest rate ve times to achieve the goals of maintain growth
and anti-deation during the second half of 2008 and 2009.
On the contrary, the interest rate adjustment is relatively smooth in the second regime, and its changes are relatively small.
During these periods, the central bank tends to carry out prudent monetary policy to promote the sustained, rapid and sound
development of the national economy. In fact, as shown in the estimation results, the central bank focuses on economic uctuation
in the second regime, and the interest rate policy is sensitive to the output gap. From Fig. 3, we can observe that the interest rate
and the output gap move in the same direction during the periods 19961997 and 20002006, reecting the positive reaction of
interest rate to output uctuation in these periods.
6. Conclusions
In this study, we consider a regime-switching forward-looking Taylor rule that nests the simple rule with Markov switching in
Zheng and Liu (2010) as a special case and realize its appropriate estimation via the two-step MLE procedure due to Kim and
Nelson (2006), Kim (2009) and Zheng and Wang (2010). Based on this regime-switching rule, we do an empirical study on the
dynamic reaction of interest rate policy to ination and output gap for China during the period from 1992Q1 to 2010Q3. Our
results showthat a two-regime forward-looking rule performs very well in modeling actual reactions of China's monetary policy. It
is also suggested that by employing this regime switching Taylor rule, we can capture signicant asymmetry in the monetary
policy reaction of the short-term interest rate to ination and output gap.
Specically, we can account for this asymmetry in the following three aspects. Firstly, this study reveals the regime switching
characteristic of monetary policy. As can be seen earlier, all empirical estimates of interest rate smoothing, policy reaction
coefcients and the variances of the interest rate shock are state-dependent. Moreover, the greater the interest rate smoothing is,
the smaller variance of the interest rate shock is, and vice visa. Secondly, the PBC behaves quite different policy reactions in
different regimes. For example, the central bank targets ination in the rst regime, i.e. the reaction of interest rate policy is mainly
focused on ination and deation, but is not so obvious on output gap. However, in the second regime, the central bank mainly
targets output gap. In this regime, the monetary policy reaction of the interest rate to ination does not follow the traditional
Taylor rule, but has an unstable framework. Finally, according to the monetary policy behavior of the central bank, the rst regime
is coincident with several major adjustments of the benchmark interest rates of lending and deposit, which plays an important role
in curbing ination and deation. On the contrary, interest rate adjustment is relative smooth in the second regime, which plays an
active role in stabilizing real output growth.
In summary, this paper, fromperspective of regime switching, has investigated the monetary policy reaction of the interest rate
to ination and output gap, and veried the asymmetric reaction of the Chinese monetary policy. We strongly believe that it
provides a new method and evidence to study China's monetary policy behavior. However, whether the monetary policy reaction
by taking interest rate as its tool is linear or nonlinear remains a lot of support from empirical studies. In particular, its dynamic
adjustment mechanism still requires some theoretical considerations. We leave this as a future research.
Acknowledgments
We would like to thank the Co-Editor Xiaodong Zhu and an anonymous referee of this Review for constructive comments and
detailed suggestions. The project was supported by the National Natural Science Foundation of China (No. 71001087), and the
Natural Science Foundation of Fujian Province of China (No. 2010J01361). All errors and omissions are solely our own.
References
Ball, L. (1999). Policy rule for open economies. In B. Taylor John (Ed.), Monetary policy rules. : University of Chicago Press.
Benassy, J. (2006). Interest rate rules, ination and the Taylor Principle: An analytical exploration. Economic Theory, 27, 143162.
Bian, Z. C. (2006). A review of empirical issues of Taylor rule and its applicability in China. Journal of Finance Research, 2006(8), 5669 (in Chinese).
Boivin, J. (2006). Has U.S. monetary policy changed? Evidence from drifting coefcients and real-time data. Journal of Money, Credit, and Banking, 38(5),
11491179.
Bruggemann, R., & Riedel, J. (2008). Nonlinear interest rate reaction functions for the UK. Working paperhttp://www.uni-graz.at/socialpolitik/papers/Brueggemann.
pdf.
Carrasco, M., Liang, H., & Ploberger, W. (2004). Optimal test for Markov switching. Working paper. : University of Rochester.
Clarida, R., Gal, J., & Gertler, M. (1998). Monetary policy rules in practice: Some international evidence. European Economics Review, 42(6), 10331068.
Clarida, R., Gal, J., & Gertler, M. (2000). Monetary policy rule and macroeconomic stability: Evidence and some theory. Quarterly Journal of Economics, 115(1),
147180.
Clarida, R., Gali, J., & Gertler, M. (2002). A simple framework for international monetary policy analysis. Journal of Monetary Economics, 49(5), 870904.
Davig, Y., & Leeper, E. (2007). Generalizing the Taylor principle. The American Economic Review, 97, 607635.
58 T. ZHENG et al. / China Economic Review 23 (2012) 4759
Di Sanzo, S. (2009). Testing for Linearity in Markov switching models: A bootstrap approach. Statistical Methods and Applications, 18, 153168.
Dolado, J., Maria-Dolores, R., & Naveria, M. (2005). Are monetary policy reaction functions asymmetric? The role of nonlinearity in the Phillips curve. European
Economic Review, 49(2), 485503.
Giannoni, M., Woodford, M., 2002, Optimal interest rate rules: II. Applications, mimeo, Princeton University, August 26.
Garcia, R. (1998). Asymptotic null distribution of the likelihood ratio test in Markov switching models. International Economic Review, 39, 763788.
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57, 357384.
Hamilton, J. D. (1996). Specication testing in Markov-switching time-series models. Journal of Econometrics, 70, 127157.
Hansen, B. E. (1992). The likelihood ratio test under nonstandard conditions: testing the Markov switching model of GNP. Journal of Applied Econometric, s, 7,
S61S82.
Hansen, B. E. (1996). Erratum: the likelihood ratio test under nonstandard conditions: Testing the Markov switching model of GNP. Journal of Applied Econometrics,
11, 195198.
Judd, J. P., & Rudebusch, G. D. (1998). Taylor's rule and the Fed: 19701997. Federal Reserve Bank of San Francisco Economic Review, 1998(3), 316.
Kim, C. -J. (1994). Dynamic linear models with Markov-switching. Journal of Econometrics, 60, 122.
Kim, C. -J. (2004). Markov-switching models with endogenous explanatory variables. Journal of Econometrics, 122, 127136.
Kim, C. -J. (2006). Time-varying parameter models with endogenous regressors. Economics Letters, 91, 2126.
Kim, C. -J. (2009). Markov-switching models with endogenous explanatory variables II: A two-step MLE procedure. Journal of Econometrics, 148, 4655.
Kim, C. -J., & Nelson, C. R. (2006). Estimation of a forward-looking monetary policy rule: A time-varying parameter model using ex post data. Journal of Monetary
Economics, 53, 19491966.
Kim, C. -J., Kishor, N. K., & Nelson, C. R. (2006). A time-varying parameter model for a forward-looking monetary policy rule based on real-time data. Working paper. :
University of Washington No UWEC-2007-32.
Kim, Y., & Kim, C. -J. (2007). Dealing with endogeneity in a time-varying parameter model: Joint estimation and two-step estimation procedures. Working Paper. :
University of Washington and university of Manitoba.
Liu, B. (2004). The designing and application of the forward-looking monetary policy rules. Journal of World Economy, 2004(4), 1218 (in Chinese).
Lu, J., & Zhong, D. (2003). Cointegration test of Taylor rule in China. Economic Research Journal, 2003(8), 7685 (in Chinese).
McCallum, B. T. (2000). Alternative monetary policy rules: A comparison with historical settings for the United States, the United Kingdom, and Japan. NBER Working
Paper 7725.
Nelson, E. (2000). UK monetary policy 19721997: A guide using Taylor rules. Bank of England Working Paper, No. 120.
Nobay, R. A., & Peel, D. A. (2003). Optimal discretionary monetary policy in a model of asymmetric Central Bank preferences. The Economic Journal, 113, 657665.
Ouyang, Z. G., & Wang, S. J. (2009). The asymmetric reaction of monetary policy to ination and real GDP in China. Economic Research Journal, 2009(9), 2738
(in Chinese).
Pagan, A. (1984). Econometric issues in the analysis of regressions with generated regressors. International Economic Review, 25(1), 221247.
Rabanal, P. (2004). Monetary policy rules and the U.S. business cycle: Evidence and implications. IMF working paper no. 164.
Ruge-Murcia, F. J. (2003). Ination targeting under asymmetric preferences. Journal of Money, Credit and Banking, 35, 763785.
Semmler, W. A., & Zhang, W. (2007). Asset price volatility and monetary policy rules: A dynamic model and empirical evidence. Economic Modelling, 24(3),
411430.
Sims, C. A. (2001). Stability and instability in US monetary policy behavior. Working paper. : Princeton University.
Sims, C. A., & Zha, T. (2006). Were there regime switching in U.S. monetary policy? The American Economic Review, 96(1), 5481.
Smith, A., Naik, P., & Tsai, C. (2006). Markov-switching model selection using KullbackLeibler divergence. Journal of Econometrics, 134, 553577.
Surico, P. (2007). The Fed's monetary policy rule and U.S. ination: The case of asymmetric preferences. Journal of Economic Dynamics and Control, 31, 305324.
Surico, P. (2007). The monetary policy of the European Central Bank. The Scandinavian Journal of Economics, 109, 115135.
Svensson, L. E. (1999). Ination Targeting as a Monetary Policy Rule. Journal of Monetary Economics, 43, 607654.
Taylor, J. (1993). Discretion versus policy rules in practice. CarnegieRochester conference series on public policy, 39, 195214.
Taylor, J. B. (1999). A historical analysis of monetary policy rules, NBER Chapters. Monetary Policy Rules (pp. 319348). : National Bureau of Economic Research,
Inc.
Taylor, J. B. (2007). Housing and monetary policy. NBER working papers 13682.
Tillmann, P. (2010). Parameter uncertainty and non-linear monetary policy rules. Macroeconomic Dynamics, Forthcoming, doi:10.1017/S1365100509991118.
Wang, S., & Zou, H. F. (2006). Taylor's rule in the condition of opening economy: To test the monetary policy of China. Statistical Research, 23(3), 4247
(in Chinese).
Xia, B., & Liao, Q. (2001). Money supply has not been appropriate for the current intermediate target of monetary policy. Economic Research Journal, 2001(8), 3343
(in Chinese).
Xie, P., & Luo, X. (2002). Taylor rule and its empirical test in China's monetary policy. Economic Research Journal, 2002(3), 312 (in Chinese).
Zhang, Y. S., & Zhang, D. Q. (2008). Interest rate rule with a role of money and its empirical test in China. Economic Research Journal, 2008(12), 6574 (in Chinese).
Zhao, J. W., & Gao, H. (2004). Robust monetary policy rules in an environment of interest rate liberalization in China. (supplement). China Economic Quarterly, 3,
4164 (in Chinese).
Zheng, T. G., & Liu, J. Q. (2010). Taylor rule with regime switching and its application to China's monetary policy. Economic Research Journal, 2010(3), 4052
(in Chinese).
Zheng, T. G., & Wang, X. (2010). An extensive study on Markov switching models with endogenous variables. Working paper.
59 T. ZHENG et al. / China Economic Review 23 (2012) 4759